May 2026 · 3 min read

Grounded, not guessed

The fastest way to lose a team's trust in a tool is to be confidently wrong once. After that, every answer gets double-checked by a human, and the tool has made work, not saved it.

Cite the source

So we do not let the system be vague. Not "this area is risky." Instead: this file, this function, this line, this owner. Something a person can verify in thirty seconds.

Check the checker

Where the stakes are high, one agent describes what the code does and a second, independent agent re-reads the code and judges that description. The second agent shares none of the first one's context, so it catches the things the first one made up.

Grounding is not a feature you add at the end. It is the difference between a demo and something a team will actually rely on.

FAQ

Common questions

It points at the exact thing: this file, this function, this line, this owner. Not "this area is risky," but a specific location a person can open and confirm in about thirty seconds. The citation is part of the answer, not a footnote you add later.

Because an agent cannot reliably catch its own mistakes — it shares the context that produced them. A second agent re-reads the source with none of the first one's assumptions, so it flags the claims the first agent invented. Independence is what makes the check worth anything.

It adds cost, so we apply it where being wrong has consequences, not to every trivial lookup. The slowdown is real but small next to the alternative: one confident error that makes a team double-check every answer by hand from then on.